Improving semantic SMT via soft semantic role label constraints on ITG alignmens

Improving Semantic SMT via Soft Semantic Role Label Constraints on ITG Alignments Meriem BELOUCIF mbeloucif@cs.ust.hk Markus SAERS masaers@cs.ust.hk Dekai WU dekai@cs.ust.hk Human Language Technology Center Hong Kong University of Science and Technology, Hong Kong Abstract We show that applying semantic role label constraints to bracketing ITG alignment to train MT systems improves the quality of MT output in comparison to the conventional BITG and GIZA alignments. Moreover, we show that applying soft constraints to SRL-constrained BITG alignment leads to a better translation system compared to using hard constraints which ap- pear too harsh to produce meaningful biparses. We leverage previous work demonstrating that BITG alignments are able to fully cover cross-lingual semantic frame alternations, by using semantic role labeling to further narrow BITG constraints, in a soft fashion that avoids losing relevant portions of the search space. SRL-based evaluation metrics like MEANT have shown that tuning towards preserving the shallow semantic structure across translations, robustly im- proves translation performance. Our approach brings the same intuition into the training phase. We show that our new alignment outperforms both conventional Moses and BITG alignment baselines in terms of the adequacy-oriented MEANT scores, while still producing comparable results in terms of edit distance metrics. 1 Introduction The quality of machine translation output relies heavily on word alignment. However, the most widespread approach to word alignment is the ad hoc method of training IBM models (Brown et al., 1990) in both directions and combining their results using var- ious heuristics. Word alignments based on inversion transduction grammars or ITGs (Wu, 1997), on the other hand, provide a more structured model leading to efficient and optimal bidirectional alignments. In this paper we introduce an improved word aligner based on applying soft seman- tic role label constraints to ITG alignment. We show that both translation adequacy and fluency can be improved by replacing the conventional GIZA++ based alignment (Och and Ney, 2000) with more semantically motivated alignments obtained through training Proceedings of MT Summit XV, vol.1: MT Researchers' Track Miami, Oct 30 - Nov 3, 2015 | p. 333 ITGs (Saers and Wu, 2009) under soft SRL constraints. The new approach is motivated by Addanki et al. (2012) who demonstrated empirically that the semantic role reorder- ings found in cross-lingual SRL frames is essentially 100% covered by ITG constraints, which suggests that it should be possible to use ITG constraints as a starting point under which to align semantic frames as we do in this paper. Our approach is further motivated by the fact that including semantic role labeling in the SMT pipeline in a different way has already been shown to increase translation quality. The semantic frame based evaluation metric MEANT, which was shown to correlate better with human adequacy judgment than conventional surface based eval- uation metrics (Lo et al., 2012), can be used as an objective function for tuning SMT. Tuning to MEANT, which attempts to optimize the degree to which a sentence’s se- mantic frames can be preserved across translation, was shown …

Content truncated to 500 words